Solar corona structures segmentation by deep learning

Segmentation of coronal holes and active regions using SDO/AIA images.

Folders and files

• data
  • ar_custom.zip - Contains files of manually annotated active regions.
  • ch_custom.zip - Contains files of manually annotated coronal holes.
  • sample_data.zip - Contains 100 193Å images + CH binary masks and 100 171Å + AR binary masks.
• notebooks
  • ar_segmentation.ipynb - Segmentation of active regions.
  • ch_segmentation.ipynb - Segmentation of coronal holes.
  • data_custom.ipynb - Downloading images and creating binary masks based on coordinates in ar_custom.zip and ch_custom.zip.
• src
  • metrics.py - Dice and IoU metrics.
  • model_scss_net.py - Segmentation model called by function scss_net.
  • prep_utils.py - Data preparation utilities.
  • utils.py - Utilities for plotting.

Libraries

List of all used libraries can be found in requirements.txt file.

• Albumentations - library used for data augmentations.
• ImageDataAugmentor - library that extends base Keras ImageDataGenerator and allows use of custom augmentations

Data

Data used for training and trained models can be found on this link.

The training data were images of the sun (SDO AIA) and a corresponding binary mask of a specific region (a coronal hole or active region).

Data sources

You can download images with HEK API or with SunPy module. Additional data sources can be found below:

Images

• https://sdo.gsfc.nasa.gov/data/
• https://suntoday.lmsal.com/suntoday/
• http://jsoc.stanford.edu/

Annotations

• CHIMERA
• Region Growth
• SPoCA

SCSS - Solar coronal structures segementation model

SCSS model is based on U-Net architecture used for segmentation of biomedical images. Summary of model can be found in modelsummary.txt file. Image below shows how SCSS model architecture.

Videos (External)

Clicking on image opens video uploaded on YouTube.